Cambricon and Hygon Lead Gains, Multiple AI Chip Manufacturers Have Adapted to DeepSeek Model

On the 22nd, several stocks in the computing power sector opened sharply higher. As of the time of writing, semiconductor foundry SMIC rose 6.29% intraday, while chip customization and semiconductor IP licensing service provider VeriSilicon Microelectronics gained 5.39%. Among AI computing power-related stocks, Hygon Information rose 17.19%, Sugon climbed 10%, and Cambricon surged 12.4%. During today’s session, Cambricon’s stock price once exceeded 1,170 yuan per share, hitting a record high, with its market capitalization surpassing 490 billion yuan.

According to the reporter’s research, some AI chip manufacturers that have not yet gone public support FP8 precision computing. Moore Threads introduced in February this year that its company, based on the MUSA Compute Capability 3.1 computing architecture, can provide native FP8 computing capabilities. Moore Threads optimized and implemented FP8 matrix multiplication on a new GPU architecture based on MUTLASS, achieving a 20%~30% performance improvement in mainstream cutting-edge large model training through FP8 mixed-precision technology.

Additionally, at the World Artificial Intelligence Conference (WAIC) in July this year, Enflame Technology launched the L600 chip, its fourth-generation chip. This chip adopts an integrated training-inference architecture and supports the FP8 precision used by the DeepSeek model during training. Metax’s Xiyun C600, launched at WAIC, also adopts an integrated training-inference solution, uses HBM3e memory, and similarly supports FP8 precision.

Previously, market analysts told the reporter that after DeepSeek was launched earlier this year, domestic computing power was adapted to DeepSeek, improving the computing power utilization rate of data centers using domestic computing power. Moreover, some domestic AI companies have begun testing with domestic chips. The reporter learned that the currently domestically trained large model based entirely on domestic computing power is iFlyTek’s Spark Model. At last week’s performance exchange meeting, Tencent’s management mentioned that the company has various chip options for inference.

Recently, research reports from several securities firms have also mentioned the domestic computing power industry chain. Tianfeng Securities’ report stated that in July, semiconductor supply chain equipment and materials saw stable growth, semiconductor foundry capacity continued to rise, and the semiconductor market remained prosperous. China Securities Co., Ltd.’s report indicated that DeepSeek’s new precision format targets the upcoming next-generation domestic chip design, suggesting that future training and inference based on the DeepSeek model are expected to increasingly utilize domestic AI chips. Additionally, Tencent’s performance meeting mentioned multiple options for inference chip supply channels. Against the backdrop of fluctuations in the international supply chain, domestic inference computing power chips are expected to provide support.